Use when packaging AISTATS code, data, proofs, simulation scripts, notebooks, random seeds, and logs as anonymous supplementary evidence or public post-acceptance artifacts, even when there is no separate artifact badge. Covers what statistically minded AISTATS reviewers inspect first and how to make Monte Carlo studies turnkey.
SkillScore60Maintained
Automatic
◆Research pack
Last verified 8/15/2026
What it changes
Passed 13 of 22 test cases with the skill, and 9 of 22 without it.
Without skill40%
With skill60%
Checks it passed that the model alone missed
The response uses the bracketed format line "[Contents]".
The response uses the bracketed format line "[Artifact role]".
The response uses the bracketed format line "[Reproduction level]".
and 1 more in the full report
11 fixed4 passed either way6 still failing1 broke
Pass-rate lift
+20 pp
Tokens
-4%
Agent turns
0%
Response time
-38%
Model gemini-3.6-flashMeasured 8/15/2026v115 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 11 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 11
Prompt
In our AISTATS submission on non-convex optimization, the proof appendix relies on L-smoothness and strong convexity assumptions, but the main theorem statement only states 'under mild regularity conditions'. Evaluate this packaging issue and output the structured evaluation block.
What was checked, without and with the skill
Checked
Without skill
With skill
The response uses the bracketed format line "[Fixes before upload]".
Missed
Passed
The response identifies omitting proof conditions from main theorem statements as a key artifact failure caught during proof inspections.
Passed
Passed
case-15 · gemini-3.6-flash · 8/15/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md65 linessingle file
AISTATS Artifact Evaluation
Use this for evidence packaging around AISTATS. The venue centers on artificial intelligence,
statistics, and machine learning, so artifacts should make statistical and computational
claims inspectable.
Artifact plan
Decide what evidence reviewers need: proof details, derivations, simulation scripts,
benchmark code, datasets, preprocessing, hyperparameter sweeps, random seeds, logs, or
qualitative examples.
Keep decision-critical evidence in the main paper or appendix; optional run files can live
in supplementary material.
Include a minimal reproduction map: environment, dependencies, hardware, commands, expected
outputs, runtime, seeds, and known nondeterminism.
For restricted data, give enough provenance and processing detail for credible
reproduction without violating data-use terms.
After acceptance, replace anonymous archives with public, licensed, citable artifacts when
feasible.
What AISTATS evidence reviewers open first
Claim type
First artifact inspected
Common failure caught
Convergence rate or regret bound
Proof appendix and constants
Condition used in the proof but missing from the theorem statement
Monte Carlo simulation
Seeded simulation script
Plots cannot be regenerated because seeds and replication counts are absent
Benchmark comparison
Training and evaluation configs
Baseline tuning budget undocumented
Bayesian or MCMC method
Sampler diagnostics and chain logs
No convergence statistics or trace evidence anywhere
Because AISTATS reviewers are often statisticians, they will rerun a small simulation far
more readily than they will retrain a deep model, so make synthetic studies turnkey before
polishing anything else.
Worked vignette: packaging a Monte Carlo study
A hypothetical submission proposes a doubly robust treatment-effect estimator with a root-n
normality guarantee, validated on synthetic causal data plus two real benchmarks.